def testMakeCovarianceUpdateOp(self):
with tf.Graph().as_default():
# Construct all arguments such that convolution kernel is applied in
# exactly one spatial location.
inputs = np.random.randn(
1, # batch_size
self.kernel_height,
self.kernel_width,
self.in_channels) # in_channels
outputs_grad = np.random.randn(
1, # batch_size
1, # output_height
1, # output_width
self.out_channels)
factor = ff.ConvDiagonalFactor((tf.constant(inputs), ),
((tf.constant(outputs_grad), ), ),
self.kernel_shape,
strides=[1, 1, 1, 1],
padding='VALID')
factor.instantiate_cov_variables()
# Completely forget initial value on first update.
cov_update_op = factor.make_covariance_update_op(0.0)
# Ensure new covariance value is same as outer-product of inputs/outputs
# vectorized, squared.
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
cov = sess.run(cov_update_op)
expected_cov = np.outer(inputs.flatten(),
outputs_grad.flatten())**2
self.assertAllClose(expected_cov, cov)
def testHasBias(self):
with tf.Graph().as_default():
inputs = tf.random_uniform(
[self.batch_size, self.height, self.width, self.in_channels])
outputs_grads = [
tf.random_uniform([
self.batch_size, self.height // self.strides[1],
self.width // self.strides[2], self.out_channels
]) for _ in range(3)
]
factor = ff.ConvDiagonalFactor((inputs, ), (outputs_grads, ),
self.kernel_shape,
self.strides,
self.padding,
data_format=self.data_format,
has_bias=True)
factor.instantiate_cov_variables()
# Ensure shape accounts for bias.
self.assertEqual([
self.kernel_height * self.kernel_width * self.in_channels + 1,
self.out_channels
], factor.cov.shape.as_list())
# Ensure update op doesn't crash.
cov_update_op = factor.make_covariance_update_op(0.0)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(cov_update_op)
def testInit(self):
with tf.Graph().as_default():
inputs = tf.random_uniform(
[self.batch_size, self.height, self.width, self.in_channels])
outputs_grads = [
tf.random_uniform([
self.batch_size, self.height // self.strides[1],
self.width // self.strides[2], self.out_channels
]) for _ in range(3)
]
factor = ff.ConvDiagonalFactor((inputs, ), (outputs_grads, ),
self.kernel_shape,
self.strides,
self.padding,
data_format=self.data_format)
factor.instantiate_cov_variables()
# Ensure covariance matrix's shape makes sense.
self.assertEqual([
self.kernel_height * self.kernel_width * self.in_channels,
self.out_channels
], factor.cov.shape.as_list())